Evolutionary Training Data Sets with n-dimensional Encoding for Neural InSAR Classifiers

Supervised training of a neural classifier and its performance not only relies on the artificial neural network (ANN) type, architecture and the training method, but also on the size and composition of the training data set (TDS). For the parallel generation of TDSs for a multi-layer perceptron (MLP) classifier we introduce evolutionary resampling and combine (erc) being based on genetic algorithms (GAs). The erc method is compared to various adaptive resample and combine techniques, namely, arc-fs, arc-lh and arc-x4. While arc methods do not consider the classifier's generalization ability, erc seeks to optimize performance by cross-validation on a validation data set (VDS). Combination of classifiers is performed by all arc methods so as to reduce the classifiers' variance, hence, erc also adopts classifier combination schemes. In order to overcome some deficiencies of the traditional approach of mapping bits of GA chromosomes to elements of a set (bit mapping) for evolution of subsets, we investigate the use of n-dimensional encoding. With this approach all available patterns are arranged in an n-dimensional space and the patterns are selected by evolving line segments conveying the data set. All algorithms are compared for a real-world problem, the classification of high resolution interferometric synthetic aperture radar (InSAR) data into several land-cover classes.
Helmut A. Mayer
Last modified: Jun 30 1998